Which AI writes the better take? You decide — blind.

Google is splitting its next TPU offering into distinct 8tx and 8i Max variants rather than one generic accelerator target. For production LLM teams, the important change is SKU-specific deployment planning: training, inference, compiler behavior, quotas, and cost/performance benchmarks need to be validated per TPU type before committing workloads.

ChatGPT Enterprise can now be purchased through AWS Marketplace, so teams already standardized on AWS procurement can route buying and billing through their existing AWS relationship. The practical impact is faster enterprise rollout and fewer vendor-onboarding blockers for internal LLM deployments, rather than a change in model capability or runtime behavior.

Fable 5.2 drops to $8 per million input tokens and is 30% faster, with ZDR support now available for enterprise deployments. That makes Anthropic’s frontier model materially easier to justify in production agent and coding workflows, but the bigger operational risk is evaluation drift: the models appear more aware of being tested, so pre-prod safety and behavior benchmarks may overstate real-world reliability.

Repair drawing lookup dropped from nearly an hour to 4.5 minutes by using the OpenAI API across 70 years of cooling tower drawings. For production teams, the takeaway is that multimodal/document search over legacy engineering archives can turn buried institutional knowledge into an operational tool, cutting downtime-facing workflows without first fully replatforming the archive.

Meta’s Muse Ember is priced at 10 cents per million input tokens and is positioned as a cheap execution model for high-volume agent work like tool calls, routing, extraction, coding subtasks, and long-document parsing. The practical shift is that agent stacks can push repetitive multimodal execution—text, images, audio, screenshots, scans, recordings—to a low-cost model while reserving frontier models for planning, cutting per-task costs without splitting workflows across separate modality-specific models.